CLASSIFICATION OF VIOLENT VIDEOS USING ENSEMBLE BOOSTING MACHINE LEARNING APPROACH WITH LOW LEVEL FEATURES

Abstract

With the fast development of monitoring video capturing gadgets to keep an eye on the human movement requires such system which perceive the suspicious, violent and dubious occasions without requiring no human intervention. Anomalous, abnormal and violent activity identification has become a prominent field of research in areas of human-computer interaction, computer vision, machine learning and digital image processing. This paper proposes a distinct technique of violent video classification, which depends upon fusion of low-level features extracted from specific frames of video since using the whole video as input may add both redundancy and noise in the learning process. We leverage low level features, and aggregate them by training a model using adaptive boosting an ensemble learning algorithm to achieve a task of Classification of Violent Videos as a binary-classification problem. The proposed technique is evaluated with three distinct datasets that are majorly used in the studies classification of violent videos. And it exhibits cutting edge classification capacity when given the task of differentiating violent and normal videos across wide variety of violent data including both crowded scenes and non-crowded scenes. The proposed work archives competitive results with an accuracy of 90.62% on hockey fight, 91.70% on movies, and 94.3% on crowd violence. The results indicates that in the view of human violence detection, our method is simple, unique and effective.